Developing a differentially private deep learning algorithm is challenging, due to the difficulty in analyzing the sensitivity of objective functions that are typically used to train deep neural networks. Many existing methods resort to the stochastic gradient descent algorithm and apply a pre-defined sensitivity to the gradients for privatizing weights. However, their slow convergence typically yields a high cumulative privacy loss. Here, we take a different route by employing the method of auxiliary coordinates, which allows us to independently update the weights per layer by optimizing a per-layer objective function. This objective function can be well approximated by a low-order Taylor's expansion, in which sensitivity analysis becomes tractable. We perturb the coefficients of the expansion for privacy, which we optimize using more advanced optimization routines than SGD for faster convergence. We empirically show that our algorithm provides a decent trained model quality under a modest privacy budget.

We propose a novel strategy for extracting features in supervised learning that can be used to construct a classifier which is more robust to small perturbations in the input space. Our method builds upon the idea of the information bottleneck by introducing an additional penalty term that encourages the Fisher information of the extracted features to be small, when parametrized by the inputs. By tuning the regularization parameter, we can explicitly trade off the opposing desiderata of robustness and accuracy when constructing a classifier. We derive the optimal solution to the robust information bottleneck when the inputs and outputs are jointly Gaussian, proving that the optimally robust features are also jointly Gaussian in that setting. Furthermore, we propose a method for optimizing a variational bound on the robust information bottleneck objective in general settings using stochastic gradient descent, which may be implemented efficiently in neural networks. Our experimental results for synthetic and real data sets show that the proposed feature extraction method indeed produces classifiers with increased robustness to perturbations.

Studies on generalization performance of machine learning algorithms under the scope of information theory suggest that compressed representations can guarantee good generalization, inspiring many compression-based regularization methods. In this paper, we introduce REVE, a new regularization scheme. Noting that compressing the representation can be sub-optimal, our first contribution is to identify a variable that is directly responsible for the final prediction. Our method aims at compressing the class conditioned entropy of this latter variable. Second, we introduce a variational upper bound on this conditional entropy term. Finally, we propose a scheme to instantiate a tractable loss that is integrated within the training procedure of the neural network and demonstrate its efficiency on different neural networks and datasets.

Application of discrete-time survival methods for continuous-time survival prediction is considered. For this purpose, a scheme for discretization of continuous-time data is proposed by considering the quantiles of the estimated event-time distribution, and, for smaller data sets, it is found to be preferable over the commonly used equidistant scheme. Furthermore, two interpolation schemes for continuous-time survival estimates are explored, both of which are shown to yield improved performance compared to the discrete-time estimates. The survival methods considered are based on the likelihood for right-censored survival data, and parameterize either the probability mass function (PMF) or the discrete-time hazard rate, both with neural networks. Through simulations and study of real-world data, the hazard rate parametrization is found to perform slightly better than the parametrization of the PMF. Inspired by these investigations, a continuous-time method is proposed by assuming that the continuous-time hazard rate is piecewise constant. The method, named PC-Hazard, is found to be highly competitive with the aforementioned methods in addition to other methods for survival prediction found in the literature.

We propose a generic confidence-based approximation that can be plugged in and simplify the auto-regressive generation process with a proved convergence. We first assume that the priors of future samples can be generated in an independently and identically distributed (i.i.d.) manner using an efficient predictor. Given the past samples and future priors, the mother AR model can post-process the priors while the accompanied confidence predictor decides whether the current sample needs a resampling or not. Thanks to the i.i.d. assumption, the post-processing can update each sample in a parallel way, which remarkably accelerates the mother model. Our experiments on different data domains including sequences and images show that the proposed method can successfully capture the complex structures of the data and generate the meaningful future samples with lower computational cost while preserving the sequential relationship of the data.

Navigating complex urban environments safely is a key to realize fully autonomous systems. Predicting future locations of vulnerable road users, such as pedestrians and cyclists, thus, has received a lot of attention in the recent years. While previous works have addressed modeling interactions with the static (obstacles) and dynamic (humans) environment agents, we address an important gap in trajectory prediction. We propose SafeCritic, a model that synergizes generative adversarial networks for generating multiple "real" trajectories with reinforcement learning to generate "safe" trajectories. The Discriminator evaluates the generated candidates on whether they are consistent with the observed inputs. The Critic network is environmentally aware to prune trajectories that are in collision or are in violation with the environment. The auto-encoding loss stabilizes training and prevents mode-collapse. We demonstrate results on two large scale data sets with a considerable improvement over state-of-the-art. We also show that the Critic is able to classify the safety of trajectories.

As a powerful tool to improve their efficiency and sustainability, most electricity systems worldwide are deploying advanced metering infrastructures to collect relevant operational data. In particular, smart meters play a key role in this transformation as they allow tracking electricity load consumption at a very disaggregated level and at high frequency rates. This data opens the possibility of developing new forecasting models with a potential positive impact in both electricity distribution and retailing activities. In this work, we present a general methodology that is able to process and forecast a large number of smart meter time series. Instead of using traditional and univariate approaches for each time series, we develop a single but complex recurrent neural network model with long short-term memory that is able to capture individual consumption patterns and also the cross-sectional relations among different households. The resulting model can accurately predict future loads (short-term) of individual consumers, even if these were not included in the original training set (out-of-sample consumers). This entails a great potential for large scale applications (Big Data) as once the single network is trained, accurate individual forecast for new consumers can be obtained at almost no computational cost. The performance of the proposed model is tested under a large set of numerical experiments by using a real world dataset with thousands of disaggregated electricity consumption time series. Furthermore, we exploit the considered dataset to explore how geo-demographic segmentation of consumers can improve the forecasting accuracy of the proposed model.

A BSTRACT We incorporate Tensor-Product Representations within the Transformer in order to better support the explicit representation of relation structure. Our Tensor-Product Transformer (TP-Transformer) sets a new state of the art on the recently-introduced Mathematics Dataset containing 56 categories of free-form math word-problems. The essential component of the model is a novel attention mechanism, called TP-Attention, which explicitly encodes the relations between each Transformer cell and the other cells from which values have been retrieved by attention. TP-Attention goes beyond linear combination of retrieved values, strengthening representation-building and resolving ambiguities introduced by multiple layers of standard attention. The TP-Transformer's attention maps give better insights into how it is capable of solving the Mathematics Dataset's challenging problems. Pretrained models and code will be made available after publication. 1 I NTRODUCTION In this paper we propose a variation of the Transformer (V aswani et al., 2017) that is designed to allow it to better incorporate structure into its representations. We test the proposal on a task where structured representations are expected to be particularly helpful: math word-problem solving, where, among other things, correctly parsing expressions and compositionally evaluating them is crucial.

We present a modern scalable reinforcement learning agent called SEED (Scalable, Efficient Deep-RL). By effectively utilizing modern accelerators, we show that it is not only possible to train on millions of frames per second but also to lower the cost of experiments compared to current methods. We achieve this with a simple architecture that features centralized inference and an optimized communication layer. SEED adopts two state of the art distributed algorithms, IMPALA/V-trace (policy gradients) and R2D2 (Q-learning), and is evaluated on Atari-57, DeepMind Lab and Google Research Football. We improve the state of the art on Football and are able to reach state of the art on Atari-57 twice as fast in wall-time. For the scenarios we consider, a 40% to 80% cost reduction for running experiments is achieved. The implementation along with experiments is open-sourced so that results can be reproduced and novel ideas tried out.

We theoretically prove that a permutation invariant property of deep neural networks largely improves its generalization performance. Learning problems with data that are invariant to permutations are frequently observed in various applications, for example, point cloud data and graph neural networks. Numerous methodologies have been developed and they achieve great performances, however, understanding a mechanism of the performance is still a developing problem. In this paper, we derive a theoretical generalization bound for invariant deep neural networks with a ReLU activation to clarify their mechanism. Consequently, our bound shows that the main term of their generalization gap is improved by n! where n is a number of permuting coordinates of data. Moreover, we prove that an approximation power of invariant deep neural networks can achieve an optimal rate, though the networks are restricted to be invariant. To achieve the results, we develop several new proof techniques such as correspondence with a fundamental domain and a scale-sensitive metric entropy. I NTRODUCTION A learning task with permutation invariant data frequently appears in various situations in data analysis. A typical example is learning on sets such as a point cloud, namely, the data are given as a set of points and permuting the points in the data does not change a result of its prediction. Another example is learning with graphs which contain a huge number of edges and nodes. Such the tasks are very common in various scientific fields [7, 8, 3], hence, numerous deep neural networks have been developed to handle such the data with invariance [15, 4, 11, 5, 13, 12].